Autrijus Tang <[EMAIL PROTECTED]> wrote:
> class Point {
> has $.x;
> has $.y;
The progress pugs makes is really impressive.
[ ... ]
> Pugs's Parrot codegen backend needs to be updated
Object attribute access, yeah. IMHO Parrot's current implementation is
wrong.
0) class construction # a bit shortened for this example
.local pmc cl, o, x, y # decls for example
cl = newclass "Point"
addattribute cl, "x" # "$x" what does Python then?
addattribute cl, "y" # "$y" what does Python then?
o = new "Point" # make instance - __init not shown
1) indexed access
$I0 = classoffset o, "Point"
x = getattribute o, $I0 # 1st attribute at ofs + 0
$I1 = $I0 + 1
y = getattribute o, $I1 # 2nd attribute at ofs + 1
The offset is cached in the bytecode, you could keep it over the whole
program life-time. If the class layout ever changes, things would break
horribly. The compiler needs intimate knowledge of the attribute order
and it's usable for an array-ish object layout only.
2) named access
x = getattribute o, "Point\0x"
This needs a full qualified attribute name "Class" ~ NUL ~ "Attribute".
That's unusable for at least Python and probably more HLLs as the
compiler has to know in which class the attribute was defined.
We should just have:
x = getattribute o, "x"
and the set equivalent:
setattribute o, "x", x
Comments welcome,
leo
